Cloud Shape Classifier 云形状分级器

home 主页 log in 登入 famous clouds 有名的云图像 about


Many people would like to see interesting clouds, but lack the spare time in which to look upwards. A Cloud Shape Classifier can help by watching the sky as people go about their routines, and showing them the best clouds at a later time. Since not everyone likes the same clouds, the Classifier learns each individual's taste, and tries to show them only the clouds that they will like.

To have the Classifier to work for you personally, you need to register, after which you can train it to like your kinds of clouds.

Verbose instructions and description


Terminological note

I have unfortunately overloaded the term "classifier". Cloud Shape Classifier is the name of the work as a whole. Small c classifier refers to a point of view that the machine holds on behalf of a user or group of users. Alternate terms would be Cloud Watching Machine and mind, respectively, but I don't like them so much.

Using a Cloud Shape Classifier

Creating a classifier

Follow a "log in" link, and type in a name and optionally a password. If you don't use a password, the classifier can be shared with other people (see shared classifiers, below).

Training your classifier

Classifiers start off in a confused state, and need to be taught to recognise good clouds. When you log in you should follow one of the links labelled "train".

You'll see it shows you four images in a grid. Click on the best image. If it is too hard to decide, you can always abstain, and another set of images will appear.

Your classifier will learn from each choice you make. To see how it is going, go to your classifier's home page. At first its opinions will fluctuate, and will often be terrible. But if you persist, it should be able to catch the drift of your taste.

Training using selected clouds

On the classifier home page, you will see that there are tick boxes next to each image. You can use these to refine your classifier. If it persistently picking an image that you despise, you can use the box, and associated "go" button, to train on a set of images including the one you want rid of. By clicking on one of the other pictures, you can quickly teach the classifier not to choose the bad one.

You can use this method on more than one image at a time. Just tick up to four boxes and any of the go buttons. You can also train your classifier using any image, if you are logged in and can find a page like this. There will be a link to train using it.

Downloading and reusing images

To download a large scale version of an image, click on any link that says "view this image", and then on the large image you see there. Most of the full-size images are between 800 and 1000 kilobytes in size. You can use these images for anything.

Shared classifiers

If you don't give your classifier a password, it becomes a public or shared classifier, and other people can help train it. The advantage of this is that your classifier can learn well with little effort on your part. On the other hand, other people might train it to like ugly clouds.

Famous clouds

Every so often a committee of classifiers is selected, and with the help of a wider poll of classifiers, selects a few clouds that they collectively think are best of all. These clouds are not directly voted for by people, only by their classifiers.

A few public classifiers are listed on the famous clouds page. You can click on these to see what those classifiers like. You then have the option of adopting that classifier, and helping to train it. Obviously, you should not only adopt classifiers that you do like, in order to give them encouragement in the direction they are already headed. Other people will have invested effort in training the classifier, and you should not disrupt their work just because they have no aesthetic understanding.

The Cloud Shape Classifier also has an interface for use in galleries. You see the same clouds as on the web, but bigger, without extraneous webisms like links and text. Gallery goers can select clouds and train a classifier by pressing special buttons.

In 2006, the Classifier was shown at ZeroOne San Jose, in California and Enjoy Public Art Gallery in Wellington. In 2007 it has been shown at Ramp in Hamilton and the Shanghai international Science and Art Exhibition.


This was made with the help of heaps of people, but I won't list many of them.

Supporting Galleries

Enjoy is an independent public art gallery based in Wellington, New Zealand.

ZeroOne San Jose (subtitled "A Global Festival of Art on the Edge"), is a festival of electronic art held in conjunction with the 13th ISEA symposium.

Ramp is a contemporary art space in Hamilton, New Zealand.

The Shanghai international Science and Art Exhibition is run by the Shanghai Science and Art Society. The New Zealand exhibit ("Geomatics and Ecomatics: Three Stories") was organised by Danny Butt.

Network support and hosting

The Cloud Shape Classifier is lucky to be hosted by CityLink, by far the best (cleverest, least evil, most useful, most interesting) telecommunications company in New Zealand.


MediaLab South Pacific lent me an office. Louise Menzies is curating the show in Wellington, and Todd Blair is doing something similar in San Jose. Lissa Mitchell is helping with setting things up.

How it works

A computer controls a camera that is pointed to the sky, and sends cloud images to the server. The server examines each image, and reduces it to 57 numbers representing various visual qualities. These numbers are stored away in a database.

Evolving Neural Networks

Each classifier has a mind in the form of a multilayer perceptron neural network. The neural network reduces the 57 numbers representing an image down to one number, which is used as a measure of goodness.

At first, the neural network has entirely arbitrary taste. When you have made some training selections, the network is altered to try to take this into account. Several copies of the network are made, and each is mutated randomly. Some of these mutated copies will give answers that better fit your selections, and some will be worse. Most of the worse ones are disposed of, and copies of the better ones are made, and then mutated again. After several iterations of this process, the best network is given back to the classifier, which uses it to pick new favourite images. This simple form of Darwinian evolution tends to be good at generalising from a few points of data, so the classifier can learn the gist of your taste.

The software is written in C and Python, and runs on Debian Linux.

Ducking under load

Because the training process uses quite a lot of resources, the server keeps track of the load it is under and trains classifiers less when the pressure gets too great. If this happens, you will see a warning message on your classifier's homepage.


Your cloud secrets are not especially safe. The cookie your browser uses to identify itself to the server consists of an SHA-1 hash of your username, password, and some other text. It doesn't change between sessions. Passwords are initially sent inplaintext and end up in the server logs. This means it is quite possible (though unlikely) for someone between you and the server to masquerade as you and train your classifier to like horrible clouds, and just a little bit harder for them to discover your password. So don't use your banking password.